gqa: ADR-0060/0062/0063/0064 unified GQA kernels + CPU cost model

Land the new GQA fused-attention kernels (ADR-0060) for prefill/decode
across long and short context, the TL discipline primitives they depend
on (ADR-0062 lazy load, ADR-0063 scratch_scope + copy_to), and the
per-op-type CPU issue cost model (ADR-0064). Remove the pre-ADR-0060
mesh-attention baseline now that the unified kernels supersede it.

ADR-0060 (long context)
- _gqa_decode.py: M-fold + 2-level chain reduce-to-root (Level-2
  intra-CUBE row-then-col + Level-1 inter-CUBE) — root-only output.
- _gqa_prefill.py: head-parallel + Ring KV rotation around C CUBEs,
  online-softmax merge per ring step, per-CUBE distributed output.
- Each merge stage wraps in scratch_scope() and persists running
  (m, l, O) via copy_to() to lift the 1 MiB scratch ceiling.

ADR-0060 §B.split.2 (short context, kv_per_cube in {1,2,4,8})
- _gqa_decode_short.py / _gqa_prefill_short.py: no cube-SP; each CUBE
  owns whole KV heads; PE-parallel heads with intra-group chain
  reduce. Prefill has no Ring KV (each head fully resident).

ADR-0062 (lazy tl.load): future-bearing TensorHandle, auto-wait at
first consuming op (dot/MATH/store/send/copy_to/composite).

ADR-0063 (tl.scratch_scope + tl.copy_to): scoped per-tile arena with
copy_to writeback primitive for persistent running state.

ADR-0064 (CPU issue cost model)
- common/cpu_issue_cost.py: per-op-type table (composite=40 ns,
  primitives=5 ns); ratios are load-bearing per D1.
- TLContext: issue_cost_table param; _emit_dispatch_overhead(kind)
  consults table with dispatch_cycles fallback (ADR-0046 §D6
  back-compat).
- Live PE_CPU paths (greenlet + legacy) construct TLContext with
  DEFAULT_CPU_ISSUE_COST so saturation lever (ADR-0060 §1) is
  measurable end-to-end.

P7 headline bench: milestone-gqa-headline writes per-panel
op_log_summary to 1H_milestone_output/gqa_headline/sweep.json. No
figure renderers yet (deferred).

Removals (pre-ADR-0060 baseline now superseded):
- benches: _attention_mesh_kv.py, _attention_mesh_mlo.py,
  _attention_mesh_mlo_2d.py, milestone_gqa_llama70b.py
- tests: test_attention_*, test_mesh_*, test_milestone_gqa_llama70b
- topology: llama70b_4sip.yaml (only consumer was the deleted diag)
- artifacts: 1H_milestone_output/gqa/ (sweep.json + 5 PNGs)
- tests/gqa/ plot helper + test (broken on Windows Tcl/Tkinter)
- ADR-0060/0061 references to deleted file paths cleaned up
  (EN + KO kept in sync).

Tests: 124/124 focused regression green (attention + Phase E + TL
discipline + triton_emu + pe_components). Full regression: 764 pass,
2 pre-existing test_bench_registry failures (stale EXPECTED_NAMES
across multiple benches, not introduced here).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
2026-06-09 18:15:59 -07:00
parent b3730a33eb
commit d282144339
52 changed files with 3952 additions and 2526 deletions
+157
View File
@@ -0,0 +1,157 @@
"""Phase 1 spec test for Phase C: long-context regression for the GQA
kernels (ADR-0060 §B "long/short context split" + ADR-0063 §A.2).
The headline panel today caps prefill at S_kv=16 / decode at S_kv≤128 —
NOT because the algorithm fails, but because the bump allocator never
recycles. ADR-0063 §A.2 (test req 3) requires a sweep at an S that
would overflow 1 MiB without recycling to complete after scope
discipline lands.
Scratch-budget estimate at C=4, P=8 (current 1 MiB pool):
decode: per-rank S_local = S_kv / 32; intermediates ≲ 500 KB at
S_kv=32K (fits today — used as regression guard).
prefill: per-rank S_local = S_kv / 4; ~16·S_local bytes per ring
step × 4 steps ≈ 64·S_local bytes. Overflows at
~64 K tokens (64 × 16K > 1 MiB).
Phase 1 (this commit): tests only — production code lands in Phase 2.
The prefill test fails today with a RuntimeError("TLContext scratch
overflow"). After Phase 2 (scratch_scope + copy_to discipline) it
completes.
"""
from __future__ import annotations
from pathlib import Path
from kernbench.benches._gqa_decode import gqa_decode_kernel # noqa: F401
from kernbench.benches._gqa_prefill import gqa_prefill_kernel # noqa: F401
from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
from kernbench.ccl.sfr_config import (
configure_sfr_intercube_multisip,
configure_sfr_intercube_ring,
)
from kernbench.policy.placement.dp import DPPolicy
from kernbench.runtime_api.bench_runner import run_bench
from kernbench.runtime_api.types import resolve_device
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import resolve_topology
TOPOLOGY_DEFAULT = Path(__file__).resolve().parents[2] / "topology.yaml"
D_HEAD = 64
DTYPE = "f16"
def _ccl_cfg():
return resolve_algorithm_config(
load_ccl_config(), name="lrab_hierarchical_allreduce",
)
def _engine_factory(t, d):
return GraphEngine(getattr(t, "topology_obj", t), enable_data=True)
# ── Decode at moderate-long context (regression guard) ───────────────
def test_decode_long_context_32k_completes():
"""Decode at S_kv=32K (C=1, P=8) — per-rank S_local=4K. Should
complete with current scratch usage (~few KB intermediates) and
must continue to complete after the rewrite.
This is a regression guard: scratch discipline shouldn't break
decode's existing long-context capability.
"""
topo = resolve_topology(str(TOPOLOGY_DEFAULT))
def _bench_fn(ctx):
configure_sfr_intercube_multisip(ctx.engine, ctx.spec, _ccl_cfg())
P = 8
S_kv = 32_768
dp_full = DPPolicy(cube="replicate", pe="replicate",
num_cubes=1, num_pes=P)
dp_kv = DPPolicy(cube="replicate", pe="row_wise",
num_cubes=1, num_pes=P)
ctx.zeros((1, 8 * D_HEAD), dtype=DTYPE, dp=dp_full, name="q_long_dec")
k = ctx.zeros((S_kv, D_HEAD),
dtype=DTYPE, dp=dp_kv, name="k_long_dec")
v = ctx.zeros((S_kv, D_HEAD),
dtype=DTYPE, dp=dp_kv, name="v_long_dec")
o = ctx.empty((1, 8 * D_HEAD),
dtype=DTYPE, dp=dp_full, name="o_long_dec")
q = ctx.zeros((1, 8 * D_HEAD),
dtype=DTYPE, dp=dp_full, name="q_long_dec_2")
ctx.launch(
"gqa_decode_long_32k",
gqa_decode_kernel,
q, k, v, o,
1, S_kv, 8, 1, D_HEAD,
1, P,
_auto_dim_remap=False,
)
result = run_bench(
topology=topo, bench_fn=_bench_fn,
device=resolve_device(None), engine_factory=_engine_factory,
)
assert result.completion.ok, (
f"decode at S_kv=32K must complete; got {result.completion}"
)
# ── Prefill at long context overflows without scope discipline ───────
def test_prefill_long_context_completes_after_scope_discipline():
"""ADR-0063 §A.2 Test Req 3: a sweep at an ``S`` that would overflow
1 MiB without recycling must complete after scope discipline lands.
With C=4 and S_kv chosen so per-rank S_local·8·4 > 1 MiB
(~16K tokens per rank ⇒ S_kv ≥ 64K), the current kernel overflows
the 1 MiB scratch pool. After Phase 2 (scratch_scope wraps each
ring step's intermediates; copy_to persists running state), it
completes.
This is the headline test that proves the S-ceiling is gone for
prefill.
"""
topo = resolve_topology(str(TOPOLOGY_DEFAULT))
def _bench_fn(ctx):
configure_sfr_intercube_ring(
ctx.engine, ctx.spec, _ccl_cfg(), ring_size=4,
)
T_q = 4
S_kv = 65_536 # 64 K — per-rank 16 K, ~2 MB scratch w/o scope
C = 4
dp_q = DPPolicy(cube="replicate", pe="replicate",
num_cubes=C, num_pes=1)
dp_kv = DPPolicy(cube="row_wise", pe="replicate",
num_cubes=C, num_pes=1)
dp_o = DPPolicy(cube="row_wise", pe="replicate",
num_cubes=C, num_pes=1)
q = ctx.zeros((T_q, D_HEAD),
dtype=DTYPE, dp=dp_q, name="q_long_pre")
k = ctx.zeros((S_kv, D_HEAD),
dtype=DTYPE, dp=dp_kv, name="k_long_pre")
v = ctx.zeros((S_kv, D_HEAD),
dtype=DTYPE, dp=dp_kv, name="v_long_pre")
o = ctx.empty((T_q * C, D_HEAD),
dtype=DTYPE, dp=dp_o, name="o_long_pre")
ctx.launch(
"gqa_prefill_long_64k",
gqa_prefill_kernel,
q, k, v, o,
T_q, S_kv, D_HEAD, C,
_auto_dim_remap=False,
)
result = run_bench(
topology=topo, bench_fn=_bench_fn,
device=resolve_device(None), engine_factory=_engine_factory,
)
assert result.completion.ok, (
f"prefill at S_kv=64K must complete after scope discipline lands; "
f"got {result.completion}"
)